#On-device

AIGoogle released EmbeddingGemma 2, a 740M-parameter multimodal embedding model under Apache 2.0 for private, on-device search and retrieval. It maps text, code, images, video, and audio into one shared space and reports a 9.92-point gain over EmbeddingGemma 1 on MTEB Code. The post lists about 191MB active RAM for quantized text-only weights and about 567MB for the full multimodal model on a Pixel 11 Pro.

AIA Chinese podcast transcript argues that Opus 5.5 makes code generation, game modding, and language migration much faster and cheaper. It cites examples such as game remakes, Photoshop clones, and ESP32 firmware projects, and says legal and open-source barriers are struggling to keep pace.

AIGoogle released Google AI Edge Foresight, a Mac app that captures meeting notes offline using the on-device EmbeddingGemma 2 model with 740 million parameters. The app offers split-screen shorthand and AI-generated notes, transcripts, and a Gemma 4-powered assistant that can answer questions from uploaded documents. Google's FAQ says it is optimized for Apple Silicon.

AIMicrosoft Copilot announced Hybrid Intelligence for Windows, which balances cloud and locally run agents to handle tasks like file wrangling and workflows on the PC. According to Satya Nadella, Copilot can use context on the PC with user permission and draw on local models when appropriate. The feature is described as coming soon.

AIAtomic Agent Desktop is a local AI-first agent available for Mac, Windows, and Linux. It offers a 4x larger context window on local models using TurboQuant, and Atomic Fusion orchestrates cloud and local models to reduce costs.

AIAtomic Agent Desktop is a free open-source app for macOS, Windows, and Linux that runs open models like Qwen and Gemma locally without an account. It connects to a cloud model only when selected, and its Fusion feature lets a cloud model plan a task while up to 8 local agents carry it out. The post's own text adds a setup wizard that checks RAM and suggests suitable models, and import from Claude Code, Codex, Hermes, and OpenClaw.

AIA ComfyUI developer generated 15-second 448×256 video in 15 seconds or less on one RTX 5090 using MiniMax H3 with FastVideo's FastH3 V2 checkpoint in four sampling steps. The setup combined sparse attention, a smaller ClipProj text encoder, a pruned INT8 checkpoint, and a fused FP4 MLP, cutting VRAM needs from 80GB to under 30GB. The custom ComfyUI node is open source.

AILiquid AI is building personal AI that runs on devices such as phones, wearables, PCs, and cars, using its Liquid Context layer, which is optimized for Snapdragon processors, to sit between models, agents, and hardware. The company's agent harness uses its own models to decide which user context to retain and how to compress it within fixed compute limits. Liquid AI is also collaborating with Mercedes-Benz Group AG to bring on-device AI to its cars and plans observability and continuous improvement loops for self-improving agents.

AIGoogle has released AI Edge Foresight, a free experimental note-taking app that transcribes meetings and audio files entirely offline on macOS. It runs on the on-device EmbeddingGemma 2 model and turns shorthand notes into polished notes based on the transcript. Google says files, meeting audio, and notes never leave the computer, and the app is currently optimized only for Macs with Apple Silicon.

AINous Research says Hermes Agent is now available on the new ASUS ProArt RTX Spark PCs, with MuseTree and ComfyUI integrations and local model support on up to 128 GB of unified memory. The post presents it as a creative stack that runs entirely on the user's own machine.

AIllama.cpp v0.6.0 adds support for multimodal Jev-like models and a performance upgrade for Metal, according to Merve Noyan. The post says models can be served with a single command, `llama serve -hf ggml-org/Clef-Flash-GGUF`, and points to a trending list of decision models on Hugging Face.

AILenovo's YOGA Pro 15 RTX Spark laptop, powered by the NVIDIA RTX Spark N1X superchip, opened blind pre-orders on October 8 at 9:00. The machine offers up to 1 Petaflop FP4 local AI compute, up to 128GB unified memory and 2TB storage, and comes preloaded with a 35B local model.

AIMicrosoft Windows and devices head Pavan Davuluri confirmed the Copilot+ PC brand has not been discontinued, saying more than 40% of commercial laptops are Copilot+ PCs shipping in tens of millions annually. He said these devices run over 2 trillion local inferences per month across search, image processing, and video calls. Microsoft plans to strengthen them through hybrid intelligence with local context, local actions, and local models.

AIGoogle's AI Edge Team open-sourced ML Drift under Apache 2.0, a GPU compute engine for on-device AI inference across OpenGL ES, OpenCL, Metal, and WebGPU. It serves as the core GPU acceleration engine within LiteRT and succeeds the legacy TFLite GPU delegate, which will no longer receive new features. The post cites benchmarks showing up to 40% lower frame latency in YouTube Shorts and up to 30% faster on-device performance in Adobe Lightroom and Photoshop.

Why it matters: The post explains how ML Drift unifies GPU shaders across platforms and replaces the TFLite GPU delegate, which matters for developers deploying on-device models.

AIAmazon has developed AI-powered smart glasses that guide delivery drivers from their van to the customer's doorstep without using their hands. Amazon applied scientist Yelin Kim will discuss the computer vision and edge AI behind the system at COLM 2026.

AIMicrosoft is adding hybrid intelligence to Copilot for Windows, letting it use local PC context and local models for tasks. Per Satya Nadella's quoted post, Windows will route each task to local or cloud models, and Copilot will act on the user's behalf only with permission.

AIMicrosoft AI says MAI-Code-1.1-Flash will run on-device inside GitHub Copilot for agentic coding. Local model calls carry no inference charge. The windowsdev post, cited as background, frames this within Windows hybrid intelligence, alongside llama.cpp and Windows ML.

AIMicrosoft is upgrading Copilot on Windows with Hybrid Intelligence, which lets it use context from the user's PC, take actions on the user's behalf, and run local models when appropriate. With the user's permission, the feature aims to add capability while stretching token usage further.

AIGeorgi Gerganov, creator of llama.cpp, said the project was showcased at a major Windows event. He credited years of community work for the software and hardware stacks now coming together, and expressed hope for wider adoption of local AI.

AINVIDIA and Microsoft announced RTX Spark laptops and compact desktops that run the full NVIDIA AI stack locally, with laptop preorders open today and sales from October 16. Microsoft also announced general availability of Microsoft Execution Containers (MXC), an OS-level infrastructure for agents to run securely in the background, while NVIDIA previewed DGX Station for Windows with 748GB of coherent memory and up to 20 petaFLOPS of FP4 compute.

Why it matters: The announcement pairs Windows agent infrastructure with local hardware, showing how agents may move onto personal computers and enterprise desktops rather than only cloud services.

AIMicrosoft says Windows will bring unmetered intelligence to PCs, letting agents work securely on-device. The post lists MAI-Code-1.1 Flash, a 137B parameter coding model with a 256K context window optimized to run on PCs, and GitHub Copilot handoffs to local models. It also describes Hybrid Intelligence, which lets Copilot act on the PC and keep sensitive work local, and Code in Copilot for building software without cloud token spend, on devices such as Surface Laptop Ultra powered by NVIDIA RTX Spark.

AIGitHub announced that Copilot will soon automatically route tasks to a local model when it is most suitable. The feature is part of Project HydraFusion and is intended to help users save on AI credits.

AILiquid AI and NVIDIA Robotics announced a collaboration around Liquid AI's models on NVIDIA hardware. The main post is brief and gives no further technical detail or figures.

AILiquid AI has released its d1 decision models, with d1-3B recommended for high-quality text and vision tasks and d1-omni-600M offered as an experimental option for smaller footprints. The weights are available for download, fine-tuning, and local deployment.

AILiquid AI's Open d1 models run across NVIDIA DGX, RTX, and Jetson hardware, with day-one llama.cpp support for deployment anywhere. Measured one request at a time, the d1-3B model's single-question latency is 8 ms on an NVIDIA RTX 4090, 16 ms on Jetson AGX Thor, 26 ms on Jetson AGX Orin 64 GB, and 50 ms on Jetson Orin Nano.

AILiquid AI built 10 live-camera demos for its d1-3B model, ranging from gesture-controlled games to content moderation, each using one forward pass per frame. In collaboration with NVIDIA Robotics, the company also showed d1-3B navigating an environment in Isaac Sim, served on a Jetson in a hardware-in-the-loop setup.

AILiquid AI has released d1-omni-600M, an experimental 600M-parameter model that handles text plus image or audio input. It combines LFM2.5-Encoder-350M with vision and audio encoders and leads the company's text benchmark comparison on toxicity detection and paraphrase identification. The post suggests uses such as voice-command routing, on-device moderation, and intent classification.

AILiquid AI released Open d1, two open-weight multimodal models in its d1 decision model family. The d1-3B model supports text and vision, while d1-omni-600M supports text plus image or text plus audio. The source says the models are meant for real-time decision making across data centers, RTX workstations, and Jetson edge devices.

AILiquid AI released two open-weight decision models, d1-3B and d1-omni-600M (experimental), built on its Liquid Foundation Models and available on Hugging Face. d1-3B scores 48.57 on the Decision Index 0.2.1, the highest among decision models under 10B parameters, and answers a question in 16 ms on an NVIDIA Jetson AGX Thor and under 50 ms on a Jetson Orin Nano. The models support text and images (d1-3B) or text with image or audio (d1-omni-600M).

AIPerplexity is open-sourcing pplx-embed-v2-late, multi-vector embedding models for text and images in one shared space, in 9B and 0.6B sizes. The 9B model can index multimodal data, the 0.6B model can run queries on device, and PDF pages can be searched without OCR. The author reports 92.4% on MADQA and 64% on BrowseComp+, with weights available on Hugging Face.

AIGoogle released the Google AI Edge Foresight app for macOS, powered by Gemma 4 and EmbeddingGemma 2. The app can connect to Google Drive to build a knowledge graph and, when transcription is active, uses local Gemma 4 E4B or Gemma 4 12B models to transcribe voice notes into new documents. EmbeddingGemma 2 is an open-weight, Apache 2 licensed 740M-parameter multimodal embedding model with an 8K context window.

AIDevelopers are running Hermes on devices such as LilyGO watches, AIPI Lite, desk gadgets, and old Android phones after the Hermes Gadget open SDK and demo were released three days ago. The post credits @NousResearch and says more boards are landing on main through contributor PRs, with the SDK available on GitHub.

AIDeepal Automobile officially announced the 2027 Deepal L06, positioned as a long-range magnetorheological AI sedan equipped with two "lobsters," namely an AI Agent and an AI Box. Chairman Deng Chenghao said the two systems let owners state their intent in a single sentence, though no launch date has been disclosed.

AIGoogle DeepMind released EmbeddingGemma 2, an open 740M-parameter embedding model that maps text, code, images, video, and audio into one 768-dimensional space. Text-only use needs a 270M-parameter footprint, about 191MB active RAM when quantized on a Pixel 11 Pro, while loading all modalities takes about 567MB. The reported MTEB Code NDCG@10 score is 78.68, about 14% above the first generation, and MTEB Multilingual v2 is 61.36, roughly flat.

AILiquid AI released two open-weight d1 decision models, d1-3B and d1-omni-600M, on Hugging Face. d1-3B scores 48.57 on the Decision Index v0.2.1 public split and answers a single question in 8 ms on an NVIDIA GeForce RTX 4090 and 50 ms on a Jetson Orin Nano. d1-omni-600M is an experimental checkpoint that handles text with images or audio and scores 15.95 on the same index.

Why it matters: The release pairs open-weight decision models with measured latency across Apple, NVIDIA, and Jetson hardware, showing how edge deployment changes what is practical.

AI@GoogleDeepMind Model page:

AIOllama announced that Google DeepMind's EmbeddingGemma 2 is now available on Ollama. The author describes it as made for consumer devices and multimodal, and gives the command ollama pull embeddinggemma-2 to download it. The quoted DeepMind post says the model is a natively multimodal open model for on-device embeddings that unifies code, images, audio, and video in a shared space.